Image Classification with Fisher Feature Analysis

Abstract

Currently, CNN-based scene classification algorithms have become mainstream. By using the features of convolutional neural networks, we propose an image classification method with Fisher feature analysis. Rich high-dimensional image descriptors can be learned through convolutional neural networks, and it is inefficient to calculate the similarity of these high feature descriptors. In order to reduce the time of feature matching and improve the accuracy of similarity descriptor matching, the algorithm adds a hidden layer between the fully-connected layer and the output layer which fine-tuning network to learn the features of low images. For solve the similarity of image feature descriptors, we use Fisher discriminant to classify images which enhance the independence between sample features. Experiments based on the Scene-15 and cifar-10 datasets show that the proposed method improves the efficiency of network feature matching and classification accuracy.